Improving Image Classification Robustness Using Predictive Data Augmentation

Abstract

Safer autonomous navigation might be challenging if there is a failure in sensing system. Robust classifier algorithm irrespective of camera position, view angles, and environmental condition of an autonomous vehicle including different size & type (Car, Bus, Truck, etc.) can safely regulate the vehicle control. As training data play a crucial role in robust classification of traffic signs, an effective augmentation technique enriching the model capacity to withstand variations in urban environment is required. In this paper, a framework to identify model weakness and targeted augmentation methodology is presented. Based on off-line behavior identification, exact limitation of a Convolutional Neural Network (CNN) model is estimated to augment only those challenge levels necessary for improved classifier robustness. Predictive Augmentation (PA) and Predictive Multiple Augmentation (PMA) methods are proposed to adapt the model based on acquired challenges with a high numerical value of confidence. We validated our framework on two different training datasets and with 5 generated test groups containing varying levels of challenge (simple to extreme). The results show impressive improvement by $\approx$ 5–20% in overall classification accuracy thereby keeping their high confidence.

Supported by Systems Safety Architecture (EPXS), Scania CV AB, Södertälje, Sweden.